APE: Further ARMv6 filter optimisations: Save 4 'ror's per round by utilising the shift feature of the 'pack halfword' instructions in the unaligned vector addition/ subtraction, better pipelining in the aligned scalarproduct(), and a new method to calculate the unaligned scalarproduct().

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@18736 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 66 insertions, 81 deletions

diff --git a/apps/codecs/demac/libdemac/vector_math16_armv6.h b/apps/codecs/demac/libdemac/vector_math16_armv6.h
index bf50d9cabd..e180429193 100644
--- a/apps/codecs/demac/libdemac/vector_math16_armv6.h
+++ b/apps/codecs/demac/libdemac/vector_math16_armv6.h
@@ -39,36 +39,33 @@ static inline void vector_add(int16_t* v1, int16_t* v2)
 
     "10:                             \n"
         "ldrh    r4, [%[v2]], #2     \n"
+        "mov     r4, r4, lsl #16     \n"
     "1:                              \n"
         "ldmia   %[v2]!, {r5-r8}     \n"
         "ldmia   %[v1],  {r0-r3}     \n"
         "mov     r5, r5, ror #16     \n"
-        "pkhbt   r4, r4, r5          \n"
+        "pkhtb   r4, r5, r4, asr #16 \n"
         "sadd16  r0, r0, r4          \n"
-        "mov     r6, r6, ror #16     \n"
-        "pkhbt   r5, r5, r6          \n"
+        "pkhbt   r5, r5, r6, lsl #16 \n"
         "sadd16  r1, r1, r5          \n"
         "mov     r7, r7, ror #16     \n"
-        "pkhbt   r6, r6, r7          \n"
+        "pkhtb   r6, r7, r6, asr #16 \n"
         "sadd16  r2, r2, r6          \n"
-        "mov     r8, r8, ror #16     \n"
-        "pkhbt   r7, r7, r8          \n"
+        "pkhbt   r7, r7, r8, lsl #16 \n"
         "sadd16  r3, r3, r7          \n"
         "stmia   %[v1]!, {r0-r3}     \n"
         "mov     r4, r8              \n"
         "ldmia   %[v2]!, {r5-r8}     \n"
         "ldmia   %[v1],  {r0-r3}     \n"
         "mov     r5, r5, ror #16     \n"
-        "pkhbt   r4, r4, r5          \n"
+        "pkhtb   r4, r5, r4, asr #16 \n"
         "sadd16  r0, r0, r4          \n"
-        "mov     r6, r6, ror #16     \n"
-        "pkhbt   r5, r5, r6          \n"
+        "pkhbt   r5, r5, r6, lsl #16 \n"
         "sadd16  r1, r1, r5          \n"
         "mov     r7, r7, ror #16     \n"
-        "pkhbt   r6, r6, r7          \n"
+        "pkhtb   r6, r7, r6, asr #16 \n"
         "sadd16  r2, r2, r6          \n"
-        "mov     r8, r8, ror #16     \n"
-        "pkhbt   r7, r7, r8          \n"
+        "pkhbt   r7, r7, r8, lsl #16 \n"
         "sadd16  r3, r3, r7          \n"
         "stmia   %[v1]!, {r0-r3}     \n"
 #if ORDER > 16
@@ -128,36 +125,33 @@ static inline void vector_sub(int16_t* v1, int16_t* v2)
 
     "10:                             \n"
         "ldrh    r4, [%[v2]], #2     \n"
+        "mov     r4, r4, lsl #16     \n"
     "1:                              \n"
         "ldmia   %[v2]!, {r5-r8}     \n"
         "ldmia   %[v1],  {r0-r3}     \n"
         "mov     r5, r5, ror #16     \n"
-        "pkhbt   r4, r4, r5          \n"
+        "pkhtb   r4, r5, r4, asr #16 \n"
         "ssub16  r0, r0, r4          \n"
-        "mov     r6, r6, ror #16     \n"
-        "pkhbt   r5, r5, r6          \n"
+        "pkhbt   r5, r5, r6, lsl #16 \n"
         "ssub16  r1, r1, r5          \n"
         "mov     r7, r7, ror #16     \n"
-        "pkhbt   r6, r6, r7          \n"
+        "pkhtb   r6, r7, r6, asr #16 \n"
         "ssub16  r2, r2, r6          \n"
-        "mov     r8, r8, ror #16     \n"
-        "pkhbt   r7, r7, r8          \n"
+        "pkhbt   r7, r7, r8, lsl #16 \n"
         "ssub16  r3, r3, r7          \n"
         "stmia   %[v1]!, {r0-r3}     \n"
         "mov     r4, r8              \n"
         "ldmia   %[v2]!, {r5-r8}     \n"
         "ldmia   %[v1],  {r0-r3}     \n"
         "mov     r5, r5, ror #16     \n"
-        "pkhbt   r4, r4, r5          \n"
+        "pkhtb   r4, r5, r4, asr #16 \n"
         "ssub16  r0, r0, r4          \n"
-        "mov     r6, r6, ror #16     \n"
-        "pkhbt   r5, r5, r6          \n"
+        "pkhbt   r5, r5, r6, lsl #16 \n"
         "ssub16  r1, r1, r5          \n"
         "mov     r7, r7, ror #16     \n"
-        "pkhbt   r6, r6, r7          \n"
+        "pkhtb   r6, r7, r6, asr #16 \n"
         "ssub16  r2, r2, r6          \n"
-        "mov     r8, r8, ror #16     \n"
-        "pkhbt   r7, r7, r8          \n"
+        "pkhbt   r7, r7, r8, lsl #16 \n"
         "ssub16  r3, r3, r7          \n"
         "stmia   %[v1]!, {r0-r3}     \n"
 #if ORDER > 16
@@ -217,80 +211,71 @@ static inline int32_t scalarproduct(int16_t* v1, int16_t* v2)
         "beq     20f                     \n"
 
     "10:                                 \n"
-        "ldrh    r2, [%[v2]], #2         \n"
-        "ldr     r0, [%[v1]], #4         \n"
-        "ldr     r3, [%[v2]], #4         \n"
-        "mov     r2, r2, lsl #16         \n"
+        "ldrh    r7, [%[v2]], #2         \n"
+        "ldmia   %[v2]!, {r4-r5}         \n"
+        "ldmia   %[v1]!, {r0-r1}         \n"
+        "mov     r7, r7, lsl #16         \n"
     "1:                                  \n"
-        "ldr     r1, [%[v1]], #4         \n"
-        "smlabt  %[res], r0, r2, %[res]  \n"
-        "ldr     r4, [%[v2]], #4         \n"
-        "smlatb  %[res], r0, r3, %[res]  \n"
-        "ldr     r0, [%[v1]], #4         \n"
-        "smlabt  %[res], r1, r3, %[res]  \n"
-        "ldr     r5, [%[v2]], #4         \n"
-        "smlatb  %[res], r1, r4, %[res]  \n"
-        "ldr     r1, [%[v1]], #4         \n"
-        "smlabt  %[res], r0, r4, %[res]  \n"
-        "ldr     r6, [%[v2]], #4         \n"
-        "smlatb  %[res], r0, r5, %[res]  \n"
-        "ldr     r0, [%[v1]], #4         \n"
-        "smlabt  %[res], r1, r5, %[res]  \n"
-        "ldr     r3, [%[v2]], #4         \n"
-        "smlatb  %[res], r1, r6, %[res]  \n"
-        "mov     r2, r6                  \n"
-        "ldr     r1, [%[v1]], #4         \n"
-        "smlabt  %[res], r0, r2, %[res]  \n"
-        "ldr     r4, [%[v2]], #4         \n"
-        "smlatb  %[res], r0, r3, %[res]  \n"
-        "ldr     r0, [%[v1]], #4         \n"
-        "smlabt  %[res], r1, r3, %[res]  \n"
-        "ldr     r5, [%[v2]], #4         \n"
-        "smlatb  %[res], r1, r4, %[res]  \n"
-        "ldr     r1, [%[v1]], #4         \n"
-        "smlabt  %[res], r0, r4, %[res]  \n"
-        "ldr     r6, [%[v2]], #4         \n"
-        "smlatb  %[res], r0, r5, %[res]  \n"
+        "pkhbt   r8, r4, r7              \n"
+        "ldmia   %[v2]!, {r6-r7}         \n"
+        "smladx  %[res], r0, r8, %[res]  \n"
+        "pkhbt   r8, r5, r4              \n"
+        "ldmia   %[v1]!, {r2-r3}         \n"
+        "smladx  %[res], r1, r8, %[res]  \n"
+        "pkhbt   r8, r6, r5              \n"
+        "ldmia   %[v2]!, {r4-r5}         \n"
+        "smladx  %[res], r2, r8, %[res]  \n"
+        "pkhbt   r8, r7, r6              \n"
+        "ldmia   %[v1]!, {r0-r1}         \n"
+        "smladx  %[res], r3, r8, %[res]  \n"
+        "pkhbt   r8, r4, r7              \n"
+        "ldmia   %[v2]!, {r6-r7}         \n"
+        "smladx  %[res], r0, r8, %[res]  \n"
+        "pkhbt   r8, r5, r4              \n"
+        "ldmia   %[v1]!, {r2-r3}         \n"
+        "smladx  %[res], r1, r8, %[res]  \n"
+        "pkhbt   r8, r6, r5              \n"
 #if ORDER > 16
         "subs    %[cnt], %[cnt], #1      \n"
-        "ldrne   r0, [%[v1]], #4         \n"
-        "smlabt  %[res], r1, r5, %[res]  \n"
-        "ldrne   r3, [%[v2]], #4         \n"
-        "smlatb  %[res], r1, r6, %[res]  \n"
-        "mov     r2, r6                  \n"
+        "ldmneia %[v2]!, {r4-r5}         \n"
+        "smladx  %[res], r2, r8, %[res]  \n"
+        "pkhbt   r8, r7, r6              \n"
+        "ldmneia %[v1]!, {r0-r1}         \n"
+        "smladx  %[res], r3, r8, %[res]  \n"
         "bne     1b                      \n"
 #else
-        "smlabt  %[res], r1, r5, %[res]  \n"
-        "smlatb  %[res], r1, r6, %[res]  \n"
+        "pkhbt   r7, r7, r6              \n"
+        "smladx  %[res], r2, r8, %[res]  \n"
+        "smladx  %[res], r3, r7, %[res]  \n"
 #endif
         "b       99f                     \n"
 
     "20:                                 \n"
         "ldmia   %[v1]!, {r0-r1}         \n"
-        "ldmia   %[v2]!, {r4-r5}         \n"
+        "ldmia   %[v2]!, {r5-r7}         \n"
     "1:                                  \n"
         "ldmia   %[v1]!, {r2-r3}         \n"
-        "smlad   %[res], r0, r4, %[res]  \n"
-        "ldmia   %[v2]!, {r6-r7}         \n"
-        "smlad   %[res], r1, r5, %[res]  \n"
-        "ldmia   %[v1]!, {r0-r1}         \n"
-        "smlad   %[res], r2, r6, %[res]  \n"
+        "smlad   %[res], r0, r5, %[res]  \n"
         "ldmia   %[v2]!, {r4-r5}         \n"
-        "smlad   %[res], r3, r7, %[res]  \n"
-        "ldmia   %[v1]!, {r2-r3}         \n"
-        "smlad   %[res], r0, r4, %[res]  \n"
+        "smlad   %[res], r1, r6, %[res]  \n"
+        "ldmia   %[v1]!, {r0-r1}         \n"
+        "smlad   %[res], r2, r7, %[res]  \n"
         "ldmia   %[v2]!, {r6-r7}         \n"
-        "smlad   %[res], r1, r5, %[res]  \n"
+        "smlad   %[res], r3, r4, %[res]  \n"
+        "ldmia   %[v1]!, {r2-r3}         \n"
+        "smlad   %[res], r0, r5, %[res]  \n"
+        "ldmia   %[v2]!, {r4-r5}         \n"
+        "smlad   %[res], r1, r6, %[res]  \n"
 #if ORDER > 16
         "subs    %[cnt], %[cnt], #1      \n"
         "ldmneia %[v1]!, {r0-r1}         \n"
-        "smlad   %[res], r2, r6, %[res]  \n"
-        "ldmneia %[v2]!, {r4-r5}         \n"
-        "smlad   %[res], r3, r7, %[res]  \n"
+        "smlad   %[res], r2, r7, %[res]  \n"
+        "ldmneia %[v2]!, {r6-r7}         \n"
+        "smlad   %[res], r3, r4, %[res]  \n"
         "bne     1b                      \n"
 #else
-        "smlad   %[res], r2, r6, %[res]  \n"
-        "smlad   %[res], r3, r7, %[res]  \n"
+        "smlad   %[res], r2, r7, %[res]  \n"
+        "smlad   %[res], r3, r4, %[res]  \n"
 #endif
 
     "99:                                 \n"
@@ -303,8 +288,8 @@ static inline int32_t scalarproduct(int16_t* v1, int16_t* v2)
         [res]"+r"(res)
         : /* inputs */
         : /* clobbers */
-        "r0", "r1", "r2", "r3",
-        "r4", "r5", "r6", "r7"
+        "r0", "r1", "r2", "r3", "r4",
+        "r5", "r6", "r7", "r8"
     );
     return res;
 }